You NEED to know about Openclaw Context Window

Context window, not model IQ, is what makes your agent turn into a "total dumbass" past a threshold.

What the context window actually is

Think of it as the agent's short-term, working memory. A viewer comment nailed it: *"the moment you go above 100K–120K context window, it feels like I'm talking to ChatGPT 3.5."* The long-term memory files (memory, soul, long-term storage) are separate and don't solve the problem — they're just reloaded into the short-term buffer every session.

It doesn't reset at midnight

The myth that OpenClaw context auto-resets to zero daily is wrong. The hosts pulled up "Jeff" on MiniMax 2.5 at the start of a fresh day and the counter was already at 105K (and later 136K) before any new conversation. Loading your life story, your girlfriend's preferences, and your car details into the bot's first wake-up fills the buffer immediately.

MiniMax silently dumps to save cost

Cheaper models "dump certain parts" of context in the background to cut compute. A call that should result in a presentation might forget whether the topic is AI or crypto videos mid-task. The hosts peg this on Chinese model providers explicitly: "for every $1 you put in, they only want to do 20 cents of work." Opus doesn't show the same drain — running on a 200K window, "Stark" stayed coherent through a full presentation while the MiniMax bot produced a boilerplate MD file and didn't even deliver it.

The fix is discipline, not settings

Don't dump your biography into the agent. Train it for one job. If it starts getting stupid, /new it — killing the session forces it to re-read only the long-term files you actually want loaded. Compaction is automatic but worse than human memory, so manual restarts beat trusting it.

Token sanity check

One Opus task burned 2.2M input tokens. That number is the receipt for why context discipline matters more than prompt quality.